A FINITENESS PROPERTY OF GRADED SEQUENCES OF IDEALS
MATTIAS JONSSON AND MIRCEA MUSTATA
o
Abstract. Given a graded sequence of ideals (a m ) m >i on X, having finite log canonical
\ threshold, we show that if there are divisors E m over X computing the log canonical
threshold of o m , and such that the log discrepancies of the divisors E m are bounded,
then the set {E m \ m > 1} is finite.
<
1. Introduction
Let X be a smooth algebraic variety over an algebraically closed field k of char-
acteristic zero. The log canonical threshold of a nonzero ideal o on X is a fundamental
invariant of the singularities of the subscheme defined by a. Originally known as the com-
plex singularity index, it shows up in many contexts related to singularities, and it has
found a plethora of applications in birational geometry (see [Kol] and [EM]).
In this note we will be interested in the behavior of this invariant in certain sequences
>
G\ ■ of ideals. Let a. = (a m ) m >i be a graded sequence of ideals on X, that is, a sequence of ideals
that satisfies ae ■ a m C ae +m for every £, m > 1. We always assume that, in addition, some
ideal a m is nonzero. The main motivating example is the graded sequence associated
to a line bundle L of nonnegative Iitaka dimension on a smooth projective variety X:
the ideal defines the base-locus of the linear system \L m \. Note that in this case the
behavior of is easy to understand if the section ring © m T(X, L m ) is finitely generated
over k. Indeed, in this case there is a positive integer p such that a mp = for all m. The
study of is useful precisely when the section ring is not finitely generated (or at least,
when this property is not known a priori).
To a graded sequence a, as above, one can associate an asymptotic version of the
log canonical threshold, by putting
let (a.) := sup m-lct(a m ).
m;a m ^(0)
This can be infinite: for example, if o. = as above, with L big, then let (a,) is infinite
if and only if L is nef (see Remark 2.2 below).
We will be concerned with the divisors that compute the log canonical thresholds of
the elements of a graded sequence. We denote by A(ord^) the log discrepancy of a divisor
Key words and phrases. Graded sequence of ideals, log canonical threshold.
2000 Mathematics Subject Classification. Primary 14F18; Secondary 14B05.
The hrst author was partially supported by NSF grants DMS-0449465 and DMS-1001740. The second
author was partially supported by NSF grant DMS-0758454 and a Packard Fellowship.
1
2
M. JONSSON AND M. MUSTATA
E over X (see §2 for the relevant definitions). The following is our main result, that gives
a positive answer to a question of Mihai Paun. 1
Theorem A. Let a, be a graded sequence of ideals on a smooth variety X such that
let (a,) < oo. If I C Z>o is a subset such that for all m G / we have a divisor E m
over X that computes lct(a m ) such that {A{oidE m ) \ m G /} is bounded, then the set
{E m | m G /} is finite.
Corollary B. Under the hypothesis in Theorem A, suppose that the set / is infinite.
Then there is a divisor E over X that computes lct(a m ) for infinitely m. In particular, E
computes let (a.).
In fact, since our proof will require replacing X by a suitable blow-up, we will need
to prove a stronger version of the above theorem, in which we replace the log canonical
threshold by the possibly higher jumping numbers, in the sense of [ELSV] (see Theorem 4.1
below for the precise statement).
Here is a sketch of the proof. Let Z m be the image of E rn on X, and let W be the
Zariski closure of \J meI Z m . We may assume that W is irreducible, and we first show that
since lct(a.) < oo, the asymptotic order of vanishing ordiy(a.) is positive. In particular,
W is a proper subset of X. If W has codimension at least two in X, then blowing-up X
along W decreases the log discrepancies of the divisors E m , and since these are bounded
above, we reduce to the case when W is a hypersurface. In this case, we use the following
result, which we believe is of independent interest.
Theorem C. Let H be a hypersurface in X, and a a nonzero ideal. Suppose that E is a
divisor over X that computes let (a). If the image Z of E on X is a proper subset of H,
and if H is smooth at the generic point of Z, then the following inequality holds
where Iz is the ideal defining Z.
Of course, as we have already mentioned, we need in fact a version of this result
that applies also to higher jumping numbers (see Theorem 3.1 below for this more general
version of the theorem). Using Theorem C, we show that if there were infinitely many
Z m that were properly contained in W, then the ideals in o. would vanish along W more
than they should. Therefore all but finitely many of the E m are equal to W (note that at
this point we are on some blow-up of our original variety).
In the following section we review some basic facts about log canonical thresholds
and higher jumping numbers. The proofs of the stronger versions of Theorems C and A
are given in §3, and respectively, §4.
M. P aim's question was motivated by the article [Siu], in which Y.-T. Siu presents part of his ar-
guments for the finite generation of the canonical ring. At the end of §6.3, he evokes a subtle point in
his approach, involving the control of an infinite sequence of blow-ups. Although expressed in a different
language, our main result shows that the infinite blow-up process in Siu's approach can be "stopped" ,
provided that the log discrepancies of the divisors computing the log canonical thresholds are bounded.
ON A FINITENESS PROPERTY OF GRADED SEQUENCES OF IDEALS
3
1.1. Acknowledgment. We are grateful to Mihai Paun for the question that led to our
main result.
2. Jumping numbers and valuations
In this section we recall some definitions and results concerning the invariants of
singularities that we will use, and set the notation for the rest of the paper. We work over
a fixed algebraically closed field k of characteristic zero. Let A be a smooth variety over
k (in particular, we assume that A is connected and separated). All ideal sheaves on X
are assumed to be coherent.
By a divisor E over X we mean a prime divisor on a normal variety Y that has a
proper birational morphism tt: Y — > X. This induces a discrete valuation of the function
field K(Y) = K(X), that we denote by ord#. As usual, we identify two such divisors
if they induce the same valuation. In particular, it follows from Hironaka's theorem on
resolution of singularities that we may assume that both Y and E are nonsingular. If we
denote by K Y /x the relative canonical divisor, then the log discrepancy of ord^ is given
by A(ords) := 1 + oy& e (K y /x)- Note that this depends on the variety X, and whenever
the variety is not clear from the context, we will write A x (ord.E)- The center of E on
X is the image cx(E) := tt(E) of E. We always consider on c x (E) the reduced scheme
structure. If a is a nonzero ideal sheaf on A, we put
ordij(a) := min{ord£;(/) | / G a ■ O x , Cx (E)} e R> -
If Z is the subscheme defined by a, we also denote this by ord^(Z).
Given an irreducible closed subset Z of A, we define the order of vanishing along Z
as follows. Consider the normalized blow-up of A along Z, and put ordz := ord^, where
E is the unique irreducible component of the exceptional divisor that dominates Z. It is
clear that in this case cx{E) = Z. Note also that ord^(a) = min^g^ ord x (a).
Let us recall the definition of multiplier ideals. For details and proofs we refer to
[Laz, §9]. Suppose that a is a nonzero ideal on A. Let \i: X' — > X be a log resolution of
(A, a), that is, ir is proper and birational, X' is nonsingular, o • Ox< = Ox'(-F) for an
effective divisor F, and F + Kx'/x has simple normal crossings. For every A G R>o ; the
multiplier ideal of a of exponent A is given by
J(a x ) :=n,O x ,(K x , /x -l\F\).
The definition is independent of the choice of log resolution.
It is clear from the above definition that if A < A', then J(a x ') C J(a x ). Further-
more, for every A there is e > such that J(<x x ) = J{of) for every t G [A, A + e]. One
says that A > is a jumping number of a if j7"(a A ) ^ J~(a x ) for every A' < A. It follows
from the definition that if we write F = a>iEi, then for every jumping number A there
is % such that Acij is an integer. In particular, the jumping numbers form a discrete set of
rational numbers.
For basic properties of the jumping numbers and applications, we refer to [ELSV].
The most important jumping number is the smallest one, known as the log canonical
4
M. JONSSON AND M. MUSTATA
threshold and denoted by let (a). This is the smallest A such that J(<x x ) ^ Ox (note that
J(a°) = O x ).
It is convenient to index the jumping numbers as follows (see [JM]). Let q be a
nonzero ideal on X. We put
lct» := min{A | q % J(a x )}.
Note that \ct° x (a) is the log canonical threshold let (a) of a. It follows from the definition
that if a 7^ O x , then riA>o«^"( aA ) = (0)' hence lct q (a) is finite. When a = O x , we make
the convention lct q (a) = oo. We will also use the notation Arn q (o) := l/lct q (o) (where
Arn stands for Arnold multiplicity). It follows from the definition that we have
. a/ N ord£;(a)
1 Arn" a = max - - " y ' -
e A(ord E ) + ord s (q)
where the maximum can be taken either over all divisors over X, or just over those lying
on a log resolution of (X, a). We say that E computes lct q (a) (or Arn q (a)) if the maximum
in (1) is achieved by E.
The most interesting of the jumping numbers is the log canonical threshold. However,
as the following lemma shows, the other jumping numbers appear naturally when we
consider higher birational models.
Proposition 2.1. Let ir: X' — » X be a proper birational morphism, with X' smooth, and
a and q nonzero ideals on X. If a' = a - Ox>, and q' = q ■ Ox>{— Kx'/x), then
let" (a) = lct q '(a').
Proof. This is an immediate consequence of (1), and of the fact that for every divisor E
over X, we have Ax(ot(1e) = Ax'(ord^) + ot(1e(Kx'/x)- D
Suppose now that a. is a graded sequence of ideals on X, and let S = {m \ (0)}.
Note that S is closed under addition. In this case we have the following asymptotic version
of the jumping numbers:
(2) lct q (o.) := sup m ■ lct q (a m ) = lim m ■ lct q (a m )
(see [JM, §2]). We put Arn q (a.) = l/lct q (a.). When q = Ox, we simply write let (a.) and
Arn(o.). Note that lct q (a.) G R >0 U {oo}. One can show that lct q (a.) = oo if and only if
lct(a.) = oo (see [JM, Corollary 6.10]).
Remark 2.2. If X is a smooth projective variety, L is a big line bundle on X, and
a. = of is the graded sequence of ideals defining the base loci of the powers of L (see
Introduction), then [ELMNP, Corollary 2.10] shows that let (a,) = oo if and only if L is
nef.
If a. is as above and E is a divisor over X, we will also consider the following
asymptotic version of the order of vanishing along E:
, , x . , ord E (o m ) ,. ordij(
ord s (o.) := mf = hm
m m^oo,meS m
ON A FINITENESS PROPERTY OF GRADED SEQUENCES OF IDEALS
5
We have the following extension of (1)
(3) Am 1 (a.) = sup ■
e A(ord E ) + ord jB (q)'
For these facts, we refer to [JM, §2]. We say that E computes lct q (a») if the supremum in
(3) is achieved by E. Note however that unlike in the case of one ideal, there may be no
divisor E that computes lct q (a.) (see [JM, Example 8.5]).
We will use the following Izumi-type estimate (see [Izu, ELS]).
Proposition 2.3. If E is a divisor over X with cx{E) = Z, then
ord^a) < ^(ord^) ■ ord z (a)
for every nonzero ideal sheaf a on X.
Proof. We may replace X by an affine open subset of the generic point of Z, and therefore
assume that X is affine. In this case we may assume that a is principal. If ord^(ci) = m,
then for a general p G Z we have ord p (a) = m. By [Kol, Lemma 8.10], there is an open
neighborhood U of p such that lct(a|f/) > l/m, and we get the assertion in the proposition
since U D Z ^ implies > let (a\ v ). □
3. AN INEQUALITY BETWEEN ORDERS OF VANISHING
We keep the notation and the conventions from §2. The following is the main result
in this section. Note that in the special case q = Ox, this recovers Theorem C in the
Introduction.
Theorem 3.1. Let H be a hyper surf ace in X, and a, q nonzero ideals on X. Suppose
that E is a divisor over X that computes lct q (a). If the center Z of E on X is a proper
subset of H , and if H is smooth at the generic point of Z , then the following inequality
holds
ot& e (Z)
(4) ord z (a) > ord H (a)
A(ord E )(l + ordtf(q))
We start by recalling a basic estimate for the log discrepancy of a valuation. For a
proof, see for example [Laz, p. 157].
Lemma 3.2. Let E be a divisor over X with Cx{E) = Z , and let £ be the generic point
of Z . If Xi, ... ,x r form a regular system of parameters of Ox,%, then
r
A(ord E ) > y^ord jE (xi).
i=i
Corollary 3.3. If H is a hypersurface in X , and E is a divisor over X such that Z :=
cx(E) is a proper subset of H , and H is smooth at the generic point of Z , then
(5) A(ord E ) > ovd E (H) + ord E (Z).
M. JONSSON AND M. MUSTATA
Proof. Let £ be the generic point of Z. Since H is smooth at £, we may choose a regular
system of parameters x\, . . . , x r of 0x,£ such that if is defined at £ by (xi). Note that by
assumption r > 2. By definition, we have ord^Z) = mm,- ord E (xj). Let i be such that
ords(xj) = oid E (Z). If i > 2, then by the lemma
A(oid E ) > ordef^i) + ord^(xj) = ord^if) + ord^(Z).
On the other hand, if i — 1, then using again the lemma we get
^4(ord E ) > ord E (xi) + ord£(x 2 ) > 2 • ord^(xi) = ord E (H) + ord E (Z).
□
Proof of Theorem 3.1. Let us put m = ord#(a) and p = ord#(q). We can write a =
O x {-H) m ■ a, and we get
(6) ord E (a) = m ■ ord E (H) + ord s (a), ord z (a) = m + ord z (a)
(note that ordz(if) = 1 since H is smooth at the generic point of Z). Since E computes
lct q (a), it follows from (1) that
ord E (a) > ordff(a) m
A{oid E ) + ord B (q) A{oid H ) + oid H (q) 1 + p
Corollary 3.3 gives ord^(iJ) < v4(ord£j) — ords(^), and combining this with (7) we deduce
/0 s , H s ord B (a) ord E (a) + m(A(ord E ) - ord E (Z)) + p ■ ord B (a)
yo) m s yi +p) • —r, — j — s j — j-t S
A(oid E ) + ord E (q) A(oid E ) + ord E (q)
ord s (a) + p ■ ord^fa) - m(ord s (q) + oid E (Z))
— JJJ -|- m
A(oid E ) + ord E (q)
Therefore ord^(a) > m(ord£(q) + oid E (Z)) — p ■ ord s (a). Using one more time the first
equation in (6), this implies
(9) (1 +p) ■ ord s (a) > m(ord£;(q) + ord^Z)) — pm ■ ord E (H).
On the other hand, by Proposition 2.3 we have ord^(a) < A(ord E ) ■ ord^(a), while clearly
orde(q) > p ■ ord E (H). Putting these together with (9) gives
(1 + p)A(oTd E ) ■ ordz(o) > (1 +p) • ordij(a)
> m(ord E (q) + ord E (Z)) — pm ■ ord E (H) > m ■ ord^Z).
Combining this with the second equality in (6), we obtain
ordg(Z)
A(ord E )(l+p).
which completes the proof of the theorem. □
Remark 3.4. In Theorem 3.1 one can replace oid E by any real valuation of K(X),
having center on X and computing lct q (a). The proof goes through if one uses the defi-
nition of A(v) from [JM, §5]. In this case, the assertion in Lemma 3.2 follows from [JM,
Corollary 5.4].
ordz(a) = m + ord^(a) > m ■ ( 1 +
ON A FINITENESS PROPERTY OF GRADED SEQUENCES OF IDEALS
7
Example 3.5. The inequality in Theorem 3.1 is optimal, at least in an asymptotic sense.
Indeed, let us consider the ideal o = x m (x,y m+1 ) in k[x,y], where m is a positive inte-
ger. Since this is a monomial ideal, one can use Howald's theorem [How] to compute its
log canonical threshold. It is easy to check that let (a) = ,™+^2 , an d this log canonical
threshold is computed by the (toric) divisor E over X = A 2 such that
ord^y^ CijX l y 3 ) = min{(m + l)i + j \ c id ^ 0}.
i,j>0
Note that Aford^) = m + 2, and the center of E on X is the origin. If we take q = Ox
and H — (x — 0), then
ordz(a) m + 1 (m+l)(m + 2)
ord.(a) (l + ^g) = - (1 + =5) = ™(™ + 3 ) '
and this converges to 1 when m goes to infinity.
Remark 3.6. Note that the right-hand side of the inequality (4) is bounded above by
ord#(a) • ^1 + i ct (j z ).(i+ or d H (q)) )> where Iz is the ideal defining Z. One could ask whether
this expression is < ord^(a), improving in this way the assertion in Theorem 3.1. However,
this is not the case: let us consider the special case m = 3 in Example 3.5, that is,
a = x 3 (x, y 4 ). With q = O x and H — (x — 0), we have ord^(a) = 4, while
ord^(a) ■ f 1 + —^--) = 3(l + ^)=?>4.
lct(Jz) / V 2 J 2
4. The main result
In this section we prove the generalized version of Theorem A in the Introduction.
We work in the same setting as in §2.
Theorem 4.1. Let o. be a graded sequence of ideals on X , and q a nonzero ideal on X
such that lct q (a.) < oo. If I C Z >0 is a subset such that for all m £ I we have a divisor
E m over X that computes lct q (a m ) such that {A(ordE m ) \ m E 1} is bounded, then the set
{E m \ me 1} is finite.
Corollary 4.2. Under the same hypothesis as in Theorem 4-1, suppose that the set I is
infinite. Then there is a divisor E over X that computes lct q (a m ) for infinitely many m.
In particular, E computes lct q (a.).
Proof of Theorem 4-1- Note that the hypothesis implies, in particular, that a m is nonzero
for every m G J. We assume that I is an infinite set, that E$ ^ Ej for all i ^ j in /
and aim to derive a contradiction. Let Z m = cx(E m ). We argue by induction on M :=
max{74(ord£;J \ i E I}. This is finite by assumption. Note that M is a positive integer,
and M = 1 if and only if all the E^s are divisors on X. At several stages in the proof we
will replace / by an infinite subset. Note that this can only decrease the value of M.
We start with the following lemma.
8 M. JONSSON AND M. MUSTATA
Lemma 4.3. With the above notation, suppose that there is an infinite subset J C I such
that W := UjizjZj is irreducible, and Zj 7^ W for all j G J. In this case
ord w (a.) > Arn(a.) > Arn q (a.) > 0.
Proof. We only need to prove the first inequality. Let C = Arn(o»), so that Arn(a m ) > Cm
for every m. If j G J, then by Proposition 2.3 we have Arn(cij) < ord^.(Oj).
We need to show that ordiy(a m ) > Cm for every m > 1. We may, of course, assume
that dm is nonzero. By hypothesis, we can find < £ < m — 1 such that the set
(10) |J Z,
j£j,j=£(mod m)
is dense in W. Since all Zj are proper subsets of W, this implies that if in (10) we
only take the union over those j G J with j = £ (mod m) and with j > N, for some
A, then the union is still dense in W. Let us fix j G J with j Q = £ (mod m), and let
C := m&x xeW ord x (Oj ) < 00 (recall that a J0 is nonzero). If mp+j Q G J, then the inclusion
C • a io C a mp+jo implies
P- ord Zmp+Jo (a m ) +ord Zmp+Jo (a io ) > ord Zmp+]Q (a mp+io ) > Arn(a mp+jo ) > C(mp + j ).
Therefore ordz mp+J0 (a m ) > Cm — — . Since we have arbitrarily large such p, and since the
union of the corresponding Z mp+ j is dense in W, we conclude that ordjy(a m ) > Cm, as
required. □
A first consequence of the lemma is that if W is the closure of Uj € /Zj, then W 7^ X.
In particular, this shows that when M = 1, we have a contradiction.
Arguing by Noetherian induction on W, we may assume that W is minimal in
X with the property that there is an infinite family of divisors (Ei) ie j as above, with
maxj^ord^J | % G /} < M. This implies first that W is irreducible. Indeed, if we
consider the irreducible decomposition W = W\ U . . . U W r , then there is j such that
Zj C Wj for infinitely many i G /. Since we may replace 7 by {i G / | Zi C Wj}, it follows
from the minimality assumption on W that W = Wj.
A second consequence of the minimality of W is that for every infinite subset J C 7,
the union Uj e jZj is dense in W. In particular, if U is an open subset of X that meets W,
then there are infinitely many i G 7 such that £/ meets Zj (and the union of these Zj n £7
is dense in IV fl £7). Therefore in order to deduce a contradiction we may replace X by U
and each o. m by its restriction to U. We may thus assume that W is nonsingular.
We claim that the induction hypothesis on M implies that W is a hypersurface in
X. Indeed, suppose that c = codim(IV, X) > 2, and let n: X' — > X be the blow-up of X
along W. If 77 is the exceptional divisor of tt, then Kx'/x — (c — 1)77. Since cx(77j) C
for every 2 G 7, it follows that Cx'(Ei) C 77, hence
A X /(ord£;J = 4x(ordjs.) - ord £i (77 X / /x ) < A x (ord E J - (c - 1).
ON A FINITENESS PROPERTY OF GRADED SEQUENCES OF IDEALS
9
If a' m = a m ■ O x > and q' = q • Ox'{— K x </x)i then by Proposition 2.1 we have lct q (a.;) =
lct q (Oj), and it follows from hypothesis and (1) that E{ computes lct q (a£) for every i E I.
Since max-fAx'tord^J | i E /} < M — 1, we have a contradiction by induction on M.
Therefore W is a smooth hypersurface in X. If Zi = W, then Ei = W, hence this
can be the case for at most one i. After discarding this i, we may assume that each Zi is
a proper subset of W. In particular, we may apply Theorem 3.1 to get
(11) ord^*) > ord w (a,) • (l + /wf^ , ^ ) •
V A(ord E J(l + ord w (q))y
Note that ordg^Zj) > 1 for all i G I. Let a = ordiy(a.). We have a > by
Lemma 4.3. Let us fix e > with e < M ^ 1+ o T d w (q)) ■ ^ we show that ord^(o m ) > am(l + e)
for every m > 1, then a = ordvy(a») > a(l + e), a contradiction. We now argue as in the
proof of Lemma 4.3. Let < £ < m — lbe such that the set in (10) is dense in W. We
fix jo ^ I such that j = £ (mod m), and let C := max^g^ ord x (Oj ). It follows from the
inclusion • Oj C a mp+ j and from (11) that for every p such that mp + j E I we have
V ■ ord Zmp+n (a rn ) > oid Zmp+n (a mp+ja ) - oid Zmp+m (a jo ) > ord W /(a mp+io )(l + e) - C.
Therefore for every such p we have ordz mp+J - (o m ) > am(l + e) — Since there are
arbitrarily large such p, and the union of the corresponding Z mp+ j is dense in W, we
conclude that ordiy(a m ) > am(l + e). As we have seen, this leads to a contradiction, and
thus completes the proof of the theorem. □
References
[ELMNP] L. Ein, R. Lazarsfeld, M. Mustata, M. Nakamaye and M. Popa, Asymptotic invariants of base
loci, Ann. Inst. Fourier (Grenoble) 56 (2006), 1701-1734. 4
[ELS] L. Ein, R. Lazarsfeld, and K. E. Smith, Uniform approximation of Abhyankar valuations in
smooth function fields, Amer. J. Math. 125 (2003), 409-440. 5
[ELSV] L. Ein, R. Lazarsfeld, K. E. Smith and D. Varolin, Jumping coefficients of multiplier ideals,
Duke Math. J. 123 (2004), 469-506. 2, 3
[EM] L. Ein and M. Mustafa, Invariants of singularities of pairs, in International Congress of
Mathematicians, Vol. II, 583-602, Eur. Math. Soc, Zurich, 2006. 1
[How] J. Howald, Multiplier ideals of monomial ideals, Trans. Amer. Math. Soc. 353 (2001), 2665-
2671. 7
[Izu] S. Izumi, A measure of integrity for local analytic algebras, Publ. RIMS Kyoto Univ. 21
(1985), 719-735. 5
[JM] M. Jonsson and M. Mustafa, Valuations and asymptotic invariants for sequences of ideals,
arXiv: 1011.3699. 4, 5, 6
[Kol] J. Kollar, Singularities of pairs, in Algebraic geometry, Santa Cruz 1995, 221-286, Proc. Symp.
Pure Math. 62, Part 1, Amer. Math. Soc., Providence, RI, 1997. 1, 5
[Laz] R. Lazarsfeld, Positivity in algebraic geometry II, Ergcbnisse der Mathematik und ihrer Grcn-
zgcbictc, 3. Folge, Vol. 49, Springer- Vcrlag, Berlin, 2004. 3, 5
[Siu] Y.-T. Siu, Techniques for the analytic proof of the finite generation of the canonical ring,
Current developments in mathematics, 2007, 177-219, Int. Press, Somerville, MA, 2009. 2
Department of Mathematics, University of Michigan, Ann Arbor, MI 48109, USA
E-mail address: mattiasj@umich.edu, mmustata@umich.edu